Musubi Cache Latents (Wan)

Second node in the jaimitoes/ComfyUI_Wan2_1_lora_trainer pipeline, right after WanDatasetConfig. Where the dataset node wrote the config, this one actually does work: it runs your images through the Wan VAE once, converts them to latent space, and saves the result so training never has to do it again. Cache the latents once, and every epoch of a multi-thousand-step run reads precomputed data off disk instead of re-encoding frames.

The pack is a GUI around kohya-ss/musubi-tuner, and this node is basically musubi's wan_cache_latents.py with a ComfyUI face. The README's whole pitch is that the entire process - write TOML, cache latents, cache text, train - runs in one shot because the nodes are chained by triggers. This is the middle of that chain.

How it works

You feed it the dataset_toml_path from WanDatasetConfig, plus a VAE. It constructs the same arguments the command-line cache script would take (via musubi's own argparse defaults, so your overrides land on top of sane defaults), then runs the cache. Output lands in the dataset1_cache_directory you set back in the config node. Because this is a long operation, the node computes an estimated batch count and drives ComfyUI's progress bar - the num_workers default scales off your CPU cores.

The inputs that matter

• vae_name - the VAE for your base model (Wan 2.1 uses wan_2.1_vae.safetensors from the Comfy-Org repackage, or Wan2.1_VAE.pth). Wrong VAE, wrong latents, bad LoRA. Get this right.
• vae_dtype - bfloat16 default; that's the right call for the balance of speed and quality.
• clip_name - None unless you're training an I2V model. This is the pack's single most repeated warning: CLIP vision exists only for image-to-video. For T2V, leave it None or the cache stage will expect a clip vision model you don't need.
• skip_existing - default off. Flip it on and the node skips files that already have cached versions, which is your resume button.
• keep_cache - on by default. Leave it on; if you turn it off, processed data can get cleaned up.
• vae_cache_cpu - off. If you're tight on VRAM on a big dataset, this shoves cached VAE outputs into system RAM instead. Slower, but it fits.

The batch_size_override and num_workers_override inputs are there if you want to diverge from the config's values; 0 means "use what the config says." device is cuda unless you enjoy waiting.

Outputs

All three wire forward:

• dataset_toml_path (STRING) - the config path, passed through untouched.
• trigger_out (any) - "latents_ok" on success, "error_trigger" on failure. Feed it to WanCacheTextEncoder's trigger_in.
• status_message (STRING) - human-readable "completed" or "Skipped: prev error" text.

Install

Same as every node in this pack - it doesn't install alone:

cd ComfyUI/custom_nodes
git clone https://github.com/jaimitoes/ComfyUI_Wan2_1_lora_trainer
..\..\..\python_embeded\python -m pip install -r requirements.txt

then restart ComfyUI. Remember this pack is Windows-portable-oriented: launch from the README's custom .bat (which loads Visual Studio Build Tools) and extract Include_and_libs.zip into python_embeded or the subprocess training bits will break.

Where people get burned

The biggest is re-caching every time you restart training. The README says it straight: you can skip the latents and text caches if you're resuming and nothing about your VAE / clip vision / text model changed. Toggle skip_existing, keep your cache folders per-LoRA (mixing them poisons later runs), and a re-run starts almost immediately instead of re-encoding your whole dataset. And if the node spits back "ERROR: VAE not found" on a name you're sure exists - check the folder the file lives in; the node resolves names against ComfyUI's vae directory specifically.